llama.cpp Releases Build b10999 with Updated Auto-Fitting Context Length
llama.cpp has released automated build b10999, featuring pull request #28849 which updates how maximum context length is calculated during auto-fitting with unified Key-Value (KV) cache. The release updates pre-compiled binaries across multiple target operating systems and backend architectures. Accurately calculating context window limits for unified KV caches helps prevent out-of-memory errors while maximizing memory utilization during local model execution. This incremental improvement helps users running long-context LLMs squeeze optimal performance out of limited GPU or system RAM. The release specifically adjusts context capacity logic under unified KV cache auto-fitting PR #28849. Pre-built binaries are provided for macOS, iOS, Windows, Android, and Linux with backends spanning CPU, CUDA 12/13, Vulkan, ROCm 10.0, OpenVINO, and SYCL.
## BACKGROUND
llama.cpp is a widely used open-source C/C++ framework for running LLMs locally across diverse hardware. The Key-Value (KV) cache stores attention vectors of previous tokens to prevent redundant computation during generation, but it quickly becomes a major memory bottleneck for long context windows. Auto-fitting context features automatically determine the maximum context length a system can hold based on available memory.